Henrik Haraldsson and Mattias Ohlsson
A New Learning Scheme for Neural Network Ensembles
We propose a new method for training an ensemble of neural networks. A population of networks is created and maintained such that more probable networks replicate and less probable networks vanish. Each individual network is updated using random weight changes. This produces a diversity among the networks which is important for the ensemble prediction using the population. The method is compared against Bayesian learning for neural networks, Bagging and a simple neural network ensemble, on three datasets. The results show that the population method can be used as an efficient neural network learning algorithm.
LU TP 02-04